Mk task queue

Task Queue & Ownership Tracker

Manages task assignment and file ownership during parallel execution.

Task Queue

Tasks are tracked in session-state/task-queue.json:

{
  "tasks": [
    {
      "id": 1,
      "description": "Implement API endpoints",
      "owner": null,
      "status": "pending",
      "ownership": ["src/api/*", "src/routes/*"],
      "blocked_by": []
    },
    {
      "id": 2,
      "description": "Implement UI components",
      "owner": null,
      "status": "pending",
      "ownership": ["src/components/*", "src/pages/*"],
      "blocked_by": []
    }
  ]
}

Task Lifecycle

pending → claimed (agent assigned) → in_progress → completed
                                   → blocked (dependency not met)

Claiming Protocol

1. Agent requests next available task
2. Queue returns lowest-ID task where: status=pending AND blocked_by all completed
3. Agent declares ownership globs
4. Queue checks for overlap with other in_progress tasks
5. If overlap: REJECT claim, report conflict to orchestrator
6. If no overlap: ASSIGN task, set status=in_progress, record owner

Ownership Enforcement

Each task declares file ownership via glob patterns. Before any file write, check if the file matches the agent's declared ownership.

Action	Owned File?	Result
Read	Any	Always allowed
Write/Edit	Owned	Allowed
Write/Edit	Not owned	STOP — report ownership violation
Write/Edit	Overlapping	STOP — report conflict to orchestrator

Integration

• Orchestrator creates the task queue when decomposing parallel work
• Parallel agents claim and complete tasks
• Orchestrator monitors queue for completion and triggers integration test
• Queue is ephemeral (session-state/) — recreated per parallel execution

Team Coordination

Team-mode coordination rules (file ownership, commit discipline, completion messages) live in .agents/skills/team-config/references/team-coordination.md, loaded by mk:team-config on team activation.

Gotchas

• Race condition on claims: Two agents reading task-queue.json simultaneously may both claim the same task. Mitigation: orchestrator is the sole claim-serializer — agents REQUEST claims through orchestrator, never self-claim directly
• Ownership globs must not overlap: src/api/* and src/api/auth/* overlap — the more specific glob must be in the SAME task, not split across agents
• Queue file doesn't auto-create: Orchestrator must create session-state/task-queue.json before dispatching parallel agents. If missing, agents should STOP and report, not create it themselves
• Completed tasks are not removed: Tasks stay in queue with status=completed for audit trail. Queue is cleaned up only when the parallel execution phase ends